Two papers published the same day on Hugging Face Daily Papers converge on a single point: diffusion transformers can become more efficient on two different fronts, video generation and robotic control.

WorldDiT (Bagel Labs) combines robot action generation and world visual modeling in a single architecture. The key point is that it does this without relying on a large pre-trained vision-language model as the backbone for action, which is the dominant expensive approach in current robotic policies. On four LIBERO benchmarks, WorldDiT sits on the Pareto frontier between parameter count and success rate, with fewer than one billion parameters.

Sol-Attn (NVIDIA Labs) accelerates inference in video generation models by making attention dynamic and sparse, without needing to retrain the model. The numbers reported on the paper page: 2 to 5 times faster end-to-end on models like Wan 2.1-14B and HunyuanVideo-13B compared to dense attention with FlashAttention-3.

For those using or building with AI, these two directions matter because they lower the cost of two of today’s most expensive tasks: generating video and controlling robots. WorldDiT reduces dependence on large, costly VLMs. Sol-Attn applies without touching model weights, so it can be inserted into existing pipelines.

As we reported on July 22, models that learn physics from video are becoming the foundation for training robotic agents without expensive simulators. WorldDiT is a concrete step in that direction.

In detail

Diffusion transformers have replaced U-Net architectures in video and image generation. The advantage is scalability: they scale better with more data and more parameters. The price is attention, the mechanism that compares every piece of input with every other piece. In videos, where each frame generates thousands of tokens, attention becomes the dominant bottleneck during inference.

Sol-Attn addresses this bottleneck with an idea: instead of computing attention over all token blocks, it computes it only on blocks that matter, and does so on the fly, without a pre-routing pass. The method is training-free: it applies to an already-trained model without changing its weights. The numbers reported on the paper page: 2.02 times faster end-to-end on Wan 2.1-14B, 2.12 times on HunyuanVideo-13B, up to 2.4 times on LTX 2.3, compared to dense FlashAttention-3. Integrated into the Sol-Engine (with kernel fusion and caching), the multipliers rise to 3.48x and 5.08x respectively. The kernel for the B200 chip is still being optimized. Code is available in the SANA repo on NVlabs.

The limitation is explicit: existing sparse attention methods struggle to be both efficient and accurate at the same time, because block selection is rigid and costly. Sol-Attn solves this by reusing proxy scores to approximate unselected blocks in a single online softmax pass. Whether this holds for long videos or complex motion remains to be verified in real use cases.

WorldDiT works on a different front. Current robotic policies chase better performance using large pre-trained VLMs as the backbone for generating actions. It’s expensive and creates dependency on models not designed for physical control. WorldDiT couples action generation with visual world modeling in a single diffusion transformer: during training, the model generates chunks of continuous actions and predicts future camera frames as normalized RGB patches. On four LIBERO simulation benchmarks, it sits on the Pareto frontier between total parameters and success rate among methods reporting all four suites, with fewer than one billion parameters.

The limitation here is clear: LIBERO is simulation, not the physical world. The jump from simulator to real robot remains open. But a sub-billion model that sits on the Pareto frontier shows that less infrastructure is needed than thought to achieve reasonable robotic control.

The two papers converge on a core idea: diffusion transformers work better when they stop treating everything the same way. WorldDiT separates action generation from the need for a huge VLM. Sol-Attn separates attention from the need to look at everything. In both cases the gain is efficiency, without sacrificing quality.